Interpretable machine learning and rolling-origin validation identify two distinct control and predictability regimes...

This study develops an interpretable machine learning-global sensitivity analysis (ML-GSA) framework to diagnose dominant process drivers in a 9-year dataset (2000-2008) from a full-scale aerated lagoon system (90,000 m³ d^-1). XGBoost surrogate models were trained for six effluent indicators and coupled with bootstrap-resolved Sobol' analysis to generate uncertainty-bounded sensitivity rankings. Predictive performance was assessed under rolling-origin validation, which revealed two distinct predictability regimes: hydraulically governed outputs generalized forward in time (TSS R² = 0.848 ± 0.05; BOD₅ R² = 0.449 ± 0.23), whereas maturation-pond outputs (pH, VSS, DO, fecal coliform) showed pronounced temporal non-stationarity, consistent with a 35% decadal increase in hydraulic loading. Sobol'-based attribution revealed a parallel two-regime control structure. Organic matter and solids removal were governed primarily by treatment stage with strong total effects (Sᵀ = 0.67 for BOD₅; Sᵀ = 0.94 for TSS), confirming hydraulic configuration as the dominant operational lever. In contrast, maturation-pond behavior was dominated by seasonal and ecological attributions. Date was the largest variance-attributing variable for VSS, and a leading driver for pH and DO, while influent VSS dominated fecal-coliform variance (Sᵀ = 0.62) through particulate shielding. Date-decomposition experiment, in which Date was replaced by Year and harmonic day-of-year terms, showed that the original Date attribution contained both long-term operational drift and within-year seasonality. Sobol' Sᵀ rankings agreed strongly with SHAP and Shapley-effect estimates (ρ ≥ 0.79), supporting robustness under input dependence. Environmentally, effluent breached BOD₅, TSS, and fecal-coliform limits in 41-87% of the record; seasonal maturation-pond disinfection was the binding constraint.